Background The relationship between cerebrospinal fluid (CSF)/serum immunoglobulin (Ig) ratios (QIg) and CSF/serum albumin ratio (Qalb) is a commonly used biomarker in routine CSF diagnostics. It is used to quantitatively determine the presence of an intrathecal Ig synthesis, which is important for the diagnosis of various inflammatory neurological diseases. Which approach to describe the relationship between Qalb and QIg should be applied has been a matter of debate for decades. Objective To compare different linear and non-linear models to determine which model fits the relationship between QIgG, QIgA, QIgM and Qalb best. Methods We included CSF and serum sample pairs collected for routine diagnostic purpose across seven sites of patients without evidence of intrathecal Ig synthesis (defined by negative oligoclonal IgG bands and the condition QIgG > QIgA > QIgM, respectively). Linear, quadratic, cubic, root and power models were fitted for the mean relationship between QIgG, QIgA, QIgM and Qalb, as well as to determine the 97.5th percentile as upper limit of normal (ULN). Model fit was evaluated by various goodness-of-fit measures including mean absolute error (MAE) and the mean percentage error (MPE). Results A total of 22,207 patients at a median age of 49 (2.5th, 97.5th percentile: 10, 82) years comprising 52% females were included. The median Qalb was 5.8 × 10−3 (2.4 × 10−3, 23 × 10−3), QIgG 3.0 × 10−3 (1.2 × 10−3, 13.4 × 10−3), QIgA 1.7 × 10−3 (0.6 × 10−3, 9.5 × 10−3) and QIgM 0.6 × 10−3 (0.2 × 10−3, 6.0 × 10−3). The goodness-of-fit measures for the mean relationship between QIgG and Qalb were similar across models [maximum difference in MAE (ΔMAE) between models was 0.01 × 10−3, in MPE (ΔMPE) 0.3 percentage points (pp)], but showed higher differences for QIgA [ΔMAE: 0.03 × 10−3, ΔMPE: 3 pp] and even higher for QIgM [ΔMAE: 0.08 × 10−3, ΔMPE: 18 pp] in favor of non-linear models. Similar findings were obtained for ULN (ΔMPE: QIgG 3 pp; QIgA 7 pp; QIgM 13 pp) again in favor of non-linear models. Discussion Linear and non-linear models can be used to approximate the relationship between QIgG and Qalb. Non-linear models explain the relationship between QIgA, QIgM and Qalb. Appropriate Qalb-dependent ULN for QIgG, QIgA, QIgM based on power model and quantile regression are provided.

The relationship between CSF/serum immunoglobulin and albumin ratio across laboratories: a retrospective, multi-center study

期刊论文
出版模式：

开放获取
发表日期：

2025年12月13日

查看全文

Harald Hegen,
Ulf Andreasson,
Juan J. Archelos-Garcia,
Gerhard Bachmaier,
Kaj Blennow,
Gabriel Bsteh,
Isabelle Hubeek,
Michael Khalil,
Jens Kuhle,
H. Bea Kuiperij,
Jan Lycke,
Igal Rosenstein,
Charlotte E. Teunissen,
Marcel M. Verbeek,
Michel Willemsen,
Henrik Zetterberg,
Janette Walde,
Florian Deisenhammer

摘要

Background

The relationship between cerebrospinal fluid (CSF)/serum immunoglobulin (Ig) ratios (Q_Ig) and CSF/serum albumin ratio (Q_alb) is a commonly used biomarker in routine CSF diagnostics. It is used to quantitatively determine the presence of an intrathecal Ig synthesis, which is important for the diagnosis of various inflammatory neurological diseases. Which approach to describe the relationship between Q_alb and Q_Ig should be applied has been a matter of debate for decades.

Objective

To compare different linear and non-linear models to determine which model fits the relationship between Q_IgG, Q_IgA, Q_IgM and Q_alb best.

Methods

We included CSF and serum sample pairs collected for routine diagnostic purpose across seven sites of patients without evidence of intrathecal Ig synthesis (defined by negative oligoclonal IgG bands and the condition Q_IgG > Q_IgA > Q_IgM, respectively). Linear, quadratic, cubic, root and power models were fitted for the mean relationship between Q_IgG, Q_IgA, Q_IgM and Q_alb, as well as to determine the 97.5^th percentile as upper limit of normal (ULN). Model fit was evaluated by various goodness-of-fit measures including mean absolute error (MAE) and the mean percentage error (MPE).

Results

A total of 22,207 patients at a median age of 49 (2.5^th, 97.5^th percentile: 10, 82) years comprising 52% females were included. The median Q_alb was 5.8 × 10⁻³ (2.4 × 10⁻³, 23 × 10⁻³), Q_IgG 3.0 × 10⁻³ (1.2 × 10⁻³, 13.4 × 10⁻³), Q_IgA 1.7 × 10⁻³ (0.6 × 10⁻³, 9.5 × 10⁻³) and Q_IgM 0.6 × 10⁻³ (0.2 × 10⁻³, 6.0 × 10⁻³). The goodness-of-fit measures for the mean relationship between Q_IgG and Q_alb were similar across models [maximum difference in MAE (ΔMAE) between models was 0.01 × 10⁻³, in MPE (ΔMPE) 0.3 percentage points (pp)], but showed higher differences for Q_IgA [ΔMAE: 0.03 × 10⁻³, ΔMPE: 3 pp] and even higher for Q_IgM [ΔMAE: 0.08 × 10⁻³, ΔMPE: 18 pp] in favor of non-linear models. Similar findings were obtained for ULN (ΔMPE: Q_IgG 3 pp; Q_IgA 7 pp; Q_IgM 13 pp) again in favor of non-linear models.

Discussion

Linear and non-linear models can be used to approximate the relationship between Q_IgG and Q_alb. Non-linear models explain the relationship between Q_IgA, Q_IgM and Q_alb. Appropriate Q_alb-dependent ULN for Q_IgG, Q_IgA, Q_IgM based on power model and quantile regression are provided.